OpenAI's sCMs Enable Real-Time AI with 50x Faster Image Generation Using 90% Less Compute

Oct 23, 2024, 05:33 PM

OpenAI has introduced simplified continuous-time consistency models (sCMs) that generate high-quality images with significantly fewer computational steps compared to traditional diffusion models. The sCMs require only two sampling steps to produce images of comparable quality to those generated by diffusion models, resulting in up to a 50x wall-clock speedup and 90% reduction in compute. The models can generate images in as little as 0.11 seconds, unlocking potential for real-time generative AI applications across various domains such as image, audio, and video generation. OpenAI researchers simplified the mathematical formulation of continuous-time consistency models, improved training stability, and scaled the models up to 1.5 billion parameters. Unlike diffusion models that require 50–100 denoising steps, consistency models convert noise directly into noise-free samples, providing a faster alternative for generating high-quality outputs.