PANews reported on October 24 that OpenAI announced that it has launched a new sCM (Simplified Consistency Model), which is based on a simplified continuous-time consistency model and provides higher training stability and scalability. It is said that sCM can generate samples of comparable quality to leading diffusion models in just two steps of sampling, greatly improving sampling speed.

Traditional diffusion models require dozens to hundreds of steps to generate samples, while sCM can be completed in just two steps by directly converting noise into noise-free samples, which is about 50 times faster. On the ImageNet 512x512 dataset, OpenAI trained sCM with 150 million parameters, and it only takes 0.11 seconds to generate a single sample. The training of sCM is based on the pre-trained diffusion model and is optimized through consistent distillation technology, which greatly reduces the computational cost while maintaining the sample quality.

OpenAI said that the progress of sCM has opened up new possibilities for efficient real-time generation of AI content and will continue to promote development in this field.