Jiaming Song, widely recognized as the father of DDIM (Denoising Diffusion Implicit Models), has announced his departure from Luma AI via a LinkedIn post. His groundbreaking work on DDIM, published in 2020, revolutionized the sampling efficiency of diffusion models, making them practical for real-world image generation. This breakthrough directly enabled the rise of popular tools such as Stable Diffusion, DALL-E, and Midjourney.
Song joined Luma AI in 2023 as Chief Scientist, leaving his position as Senior Research Scientist at NVIDIA. At Luma, he oversaw critical technical shifts: from 3D generation (Genie series) to video generation (Dream Machine) and, most recently, to multimodal foundation models (Uni-1.1). These pivots mirrored the broader AI industry's evolution over the past three years.
During his tenure, Luma AI secured Series B funding led by a16z and built systems that many considered beyond the reach of a startup. Song noted that he had the opportunity to work with brilliant researchers and help build systems that were previously thought impossible at a company of that scale.
Song's academic background includes a B.S. in Computer Science from Tsinghua University (top 1%) and a Ph.D. from Stanford under Stefano Ermon. He also held research roles at Megvii, OpenAI, Meta FAIR, and NVIDIA. His seminal paper on DDIM, along with earlier work like SDEdit, established him as a pioneer in diffusion model acceleration.
Now, after nearly three years at Luma, Song has updated his LinkedIn profile to reflect his departure. He has not yet revealed his next position or any entrepreneurial plans. The AI community will be watching closely to see where this key figure lands next.