Follow us on Google News (click on ☆)
Thanks to a new model, the AI produces images and videos 50 times faster, without compromising on quality. The key? A two-step system.
Traditional diffusion models, widely used for generating images and videos, operate with hundreds of steps of progressive noise reduction. This process, although effective, requires significant time and computational resources.
To address this, Cheng Lu and Yang Song from OpenAI developed a "continuous-time consistency model" (sCM) that simplifies the creation process. With just two steps, this model can generate high-quality samples, surpassing the performance of traditional diffusion methods.
With over 1.5 billion parameters, OpenAI's sCM operates at record speed: 0.11 seconds to produce an image on an A100 processor (for reference). In comparison, previous models required several seconds and significantly more power.
This speed could revolutionize the use of AI in domains requiring real-time generation. The image, video, and audio sectors are poised to benefit from this breakthrough.
Tested on dense datasets, the sCM maintains image quality that rivals diffusion models. Its FID score (Fréchet Inception Distance, a measure used to assess the quality of images created by a generative model) is only 10% lower than the best models, which speaks to its efficacy.
By limiting the required computational resources, the sCM also becomes a more environmentally friendly and economical solution compared to current models, although, in contrast, this should contribute to its wider adoption, and thus overall energy consumption... OpenAI hopes to further optimize its speed for demanding industrial applications.
Ultimately, the sCM could provide a technological foundation for ultra-fast generative AIs. This leap forward hints at a future where AI will produce media with unprecedented quality and fluidity, and undoubtedly real-time videos.