A leading tech company has unveiled “consistency models” for generating images, audio, and video, promising faster and higher-quality outputs compared to traditional diffusion models. These models use direct noise-to-data mapping for rapid one-step creation, while also allowing multi-step sampling for quality vs. computational cost balancing. They excel in data editing tasks without specific training. Consistency models can be trained either via diffusion model distillation or as standalone generative models. They outperformed previous models in one-step and few-step sampling, achieving record FID scores on CIFAR-10 and ImageNet 64×64. Pioneered by AI experts Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever, these advancements point towards a transformative future in high-quality content generation.

OpenAI introduced consistency models, a new generative model family enhancing the speed and quality of image, audio, and video sample generation. These models can produce samples in a single step by mapping noise directly to data, offering an advantageous alternative to traditional diffusion models. They also support multi-step sampling and exhibit remarkable zero-shot data editing capabilities for tasks like image inpainting and super-resolution. Consistency models, trained via distillation from pre-trained diffusion models or as standalone models, outperform existing methods, achieving state-of-the-art FID scores on benchmarks like CIFAR-10 and ImageNet 64×64. This innovation, spearheaded by researchers including Yang Song and Ilya Sutskever, sets new standards in generative modeling.

OpenAI has unveiled “consistency models,” a new class of generative models that generate high-quality images, audio, and video faster than traditional diffusion models. These models map noise directly to data for rapid one-step generation and also support multi-step sampling to enhance sample quality. They excel in zero-shot data editing—tasks like image inpainting and colorization—without specific training. Consistency models can be trained independently or by distilling pre-trained diffusion models, achieving state-of-the-art FID scores. They outperform existing techniques and rival one-step, non-adversarial models, signaling a new benchmark in generative AI. Authors include Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever.

OpenAI has unveiled its advanced Voice Engine, a text-to-speech model that creates highly realistic human-like audio from a mere 15-second voice sample. Developed since late 2022 and tested internally, the model generates speech by learning from paired audio-text data and using a diffusion process to replicate vocal nuances. Publicly showcased with ChatGPT’s Voice Mode in September 2023 and later released via a TTS API, Voice Engine prioritizes ethical deployment. Safety measures include watermarking, monitoring, and stringent usage policies. As GPT-4o introduces native audio capabilities, OpenAI remains committed to balancing innovation with comprehensive risk mitigation.

OpenAI and News Corp have entered a multi-year global partnership, granting OpenAI access to News Corp’s premium journalism, including content from The Wall Street Journal, Barron’s, MarketWatch, and The Times. This collaboration aims to enhance AI responses with reliable news sources and uphold high journalistic standards. News Corp will share its expertise to maintain accuracy and integrity in OpenAI’s offerings. Both Robert Thomson, News Corp’s CEO, and Sam Altman, CEO of OpenAI, emphasized the historic significance of this union, aiming to set new benchmarks for digital-era journalism and technological integration.