OpenAI pushes the boundaries of AI image and video creation. Now, media generation happens in a flash, making the wait time almost invisible.
Thanks to a new model, AI produces images and videos 50 times faster, without compromising on quality. The key? A two-step system.
Traditional streaming models, widely used to generate images and videos, operate with hundreds of progressive denoising steps. This process, although efficient, requires time and high computational resources.
To address this, Cheng Lu and Yang Song of OpenAI developed a “continuous-time consistency model” (sCM) to simplify the creation process. In just two steps, this model can generate high-quality samples, outperforming conventional diffusion methods.
With more than 1.5 billion parameters, OpenAI’s sCM operates in record time: 0.11 seconds to produce an image on a processor A100 (as an example). In comparison, earlier models require several seconds and much more power.
This speed could transform the use of AI in areas requiring real-time generation. The image, video and audio sectors are therefore poised to benefit from this progress.
Tested on dense datasets, sCM maintains image quality rivaling diffusion models. Its FID score (Fréchet Inception Distance, a measure used to evaluate the quality of images created by a generative model) is thus only 10% lower than that of the best models, which testifies to its effectiveness.
By limiting the necessary computing resources, sCM also becomes a more ecological and economical solution than current models, although in return this should contribute to its democratization, and therefore overall energy consumption… OpenAI hopes to further optimize its speed for demanding industrial applications.
Ultimately, sCM could provide a technological basis for ultra-fast generative AI. This leap forward suggests a future where AI will produce media with unprecedented quality and fluidity, and undoubtedly videos in real time.
Related News :