OpenAI's TrigFlow Achieves 2-Step Image Generation in 0.11 Seconds, 90% Less Compute, 1.5B Parameters

Oct 27, 2024, 05:11 PM

A recent paper from OpenAI details advancements in generative modeling through a new framework called TrigFlow. This innovative approach enables two-step image generation that achieves quality comparable to traditional hundred-step diffusion models while using 90% less computational resources. The model can generate images in just 0.11 seconds, significantly enhancing efficiency without compromising on quality. Additionally, the framework scales continuous-time consistency models to 1.5 billion parameters, allowing for image quality that is 50 times faster than previous methods. These developments mark a significant leap in the capabilities of text-to-image synthesis, bridging the gap between masked image modeling and diffusion techniques.