Diffusion models are effective, but they are slow which isn't practical for prolonged use or app implementation.
OpenAI shared a new approach which they call "sCM" that is a whopping 50 times faster than the standard diffusion model and creates a comparable result in just 2 sampling steps.
An example OpenAI gave was an image at 512x512 resolution that sCM generated in just 11ms, which took diffusion model a little over 6s.
Why is this a big deal?
sCM approach is a big deal because it can be used for practical purposes within apps that need to dynamically create an image. We are already at a very good-looking image quality for smaller resolutions that come in fraction of a second, it will only get better over time.
This continuous-time consistency model is close to the performance of the original diffusion model in terms of quality, but converts noise directly into an image as opposed to using dosens of steps for a single image, which significantly increases the speed of generation.
If you'd like to see some examples and a detailed explanation of this new approach to generative AI, you can read more at
OpenAI's official paper.
