Jiaming Song

Research

Song completed a PhD in computer science at Stanford and stayed on as a postdoc with Stefano Ermon. With Chenlin Meng and Ermon, Song wrote the 2020 DDIM paper, which introduced an accelerated sampler for diffusion models, and Song co-authored follow-up work on guided image synthesis and editing with SDEdit and on solving image restoration problems with Denoising Diffusion Restoration Models.¹²⁵⁶

In 2025 Song, Linqi Zhou and Ermon introduced Inductive Moment Matching, a class of generative models for one- or few-step sampling trained in a single stage without distillation. In a May 2026 note, Song described how autoregressive and diffusion methods can coexist in one generative system and discussed how flow maps relate to inference-time scaling.⁷⁸

Career

After Stanford, Song was a research scientist at NVIDIA Research, working on multimodal generation and foundation model projects. As Chief Scientist at Luma AI, Song led research through the company's shifts from 3D generation with Genie to the Ray family of video generation models and then to Uni-1, a multimodal reasoning model for image generation and editing, working across model architectures, training infrastructure and data processing.¹

Work and contributions

• DDIM (Denoising Diffusion Implicit Models) Open-source code

  First author, with Chenlin Meng and Stefano Ermon, of the 2020 DDIM paper. The ermongroup/ddim repository, under an MIT License with Song as copyright holder, implements the sampler and notes that DDIM is also available in Hugging Face Diffusers through DDIMPipeline.²³⁴¹

• Luma AI models (Genie, Ray, Uni-1) Proprietary product

  Led research at Luma AI through its 3D generation (Genie), video generation (Ray), and unified multimodal (Uni-1) models, working across model architectures, training infrastructure, and data processing, per Song's homepage.¹

Timeline

1. Oct 6, 2020Released "Denoising Diffusion Implicit Models" (DDIM) with Chenlin Meng and Stefano Ermon.²
2. Aug 2, 2021Co-authored SDEdit, a method for guided image synthesis and editing with stochastic differential equations.⁵
3. Jan 27, 2022Co-authored Denoising Diffusion Restoration Models for image restoration tasks.⁶
4. Mar 10, 2025Co-authored Inductive Moment Matching, a single-stage approach to few-step generative sampling.⁷
5. May 31, 2026Published a note on inference-time scaling for generative pre-training.⁸

Documented roles

The sources do not confirm a current role; these are the roles they document.

• Chief Scientist, Luma AI¹
• Research scientist, NVIDIA Research¹
• Postdoc with Stefano Ermon, Stanford University¹

Companies in their history

• Luma AI 2 people in the wiki
• NVIDIA 43 people in the wiki
• Stanford University 40 people in the wiki

Related people

• Amit JainConnected to Luma AI
• Wei-Ning HsuMultimodal models · Image, video, and 3D generation
• Saining XieImage, video, and 3D generation · Multimodal models
• Alexandre DéfossezImage, video, and 3D generation · Multimodal models
• Nataniel RuizImage, video, and 3D generation · Multimodal models
• Xu TanImage, video, and 3D generation · Multimodal models
• Aditya RameshImage, video, and 3D generation · Multimodal models
• Sanja FidlerImage, video, and 3D generation
• Shruthan RadhakrishnaMultimodal models
• Tero KarrasImage, video, and 3D generation