Jiaming SONG

During his time at Stanford, Jiaming Song made significant strides in the development of diffusion models, culminating in the creation of Denoising Diffusion Implicit Models (DDIM). It is the first work that accelerates diffusion models by up to 50× while maintaining the ability to produce diverse, high-fidelity samples, representing a major leap forward in the field of diffusion models and generative AI.

DDIM played a key role in making diffusion models more efficient and widely usable by the community. The work has been adopted in various notable research projects and production-level systems such as DALL-E 2, Imagen, and Stable Diffusion.

Jiaming is currently the Chief Scientist at Luma AI, allowing him to translate these theoretical advancements into practical, cutting-edge applications. He led the development of Dream Machine, capable of generating complex, coherent video sequences from text prompts, image conditions, as well as other control signals.

Dream Machine pushes the boundaries of what is possible with large-scale AI models, incorporating insights from his earlier work on diffusion models and extending them to the challenging domain of video generation.

In terms of temporal consistency, visual quality, and camera movement, Dream Machine is setting new standards in the field of generative AI for video. The ability to generate high-quality, coherent video content through simple prompts is expected to fundamentally change the way people create and consume content in fields such as entertainment, education, marketing, and communication.

Dream Machine has seen tremendous growth as well, attracting over 1 million users 4 days since its release and releasing new features such as extend, keyframes, loop, and end frame. These innovations open up new possibilities for AI to assist humans in content creation.