A new generation of AI image models is redefining where and how images are created. Researchers have developed a system that produces high-quality visuals using roughly 10 times fewer processing steps than conventional diffusion models, dramatically improving efficiency and speed, tells Live Science.
Most modern text-to-image tools rely on diffusion, a process that gradually transforms random noise into a coherent image over 30–50 iterative steps. While effective, this method is computationally intensive and typically requires powerful cloud-based GPUs. The reliance on remote servers introduces latency, raises energy demands, and requires user data to be transmitted off-device.
The new model, called Stable Diffusion 3.5 Flash, compresses this process into just a handful of steps, reportedly as few as four. Instead of refining images incrementally, it learns to make larger, more efficient jumps during generation while preserving visual quality. This reduction in steps significantly lowers computational requirements, enabling the model to run on consumer hardware such as smartphones and laptops.
This shift has broader implications beyond speed. Running AI locally enhances privacy by keeping user prompts and generated content on-device. It also reduces dependence on energy-intensive data centers, addressing growing concerns about AI’s environmental footprint.
Industry adoption is already underway. The model has been licensed for integration into upcoming on-device AI platforms, signaling a move toward embedding generative capabilities directly into everyday computing devices.
By collapsing complex diffusion pipelines into lightweight workflows, this approach marks a turning point for generative AI. It opens the door to real-time image creation, broader accessibility, and a more decentralized AI ecosystem, where powerful tools operate directly in users’ hands rather than distant servers.