Jiaming Song (@baaadas) / X

Jiaming Song

874 posts

Jiaming Song

@baaadas

I need a vacation

Joined November 2014

Pinned
Jiaming Song
@baaadas
May 31
time for some vacation stuff
00:00
Jiaming Song
@baaadas
May 31
Yesterday was my last day at @LumaLabsAI. Over the last three years, I had the privilege of helping drive the company's transition from 3D AI to video generation and native multimodal foundation models. I am grateful to have worked alongside an extraordinary group of
24K
Jiaming Song
@baaadas
Mar 11, 2025
As one of the people who popularized the field of diffusion models, I am excited to share something that might be the “beginning of the end” of it. IMM has a single stable training stage, a single objective, and a single network — all are what make diffusion so popular today.
156K
Jiaming Song
@baaadas
Mar 13, 2022
"The paper is not novel because some arxiv paper in February of 2022 already did it" -- I recall that the #icml2022 submission deadline was on January of 2022? I am fine with the paper getting rejected, but not all of us have time machines 🤣 @icmlconf
Jiaming Song
@baaadas
Mar 31, 2025
Found this figure while randomly reading @torchcompiled 's blog. It reminds me about something I had thought about around 4.5 years ago: adding noise is like drawing balls in persistent homology. Maybe we can use it to analyze the optimal way to sample timesteps during training?
34K
Jiaming Song
@baaadas
Nov 22, 2022
📢 We are looking for highly motivated #ML #AI Ph.D. students to work with us at NVIDIA Research as summer #interns next year. We encourage applicants with experience in generative modeling in one of these domains:
Jiaming Song
@baaadas
May 20, 2023
@chenlin_meng, @ArashVahdat, and I are presenting the #diffusion model tutorial at #CVPR2023 on June 18 (…3-tutorial-diffusion-models.github.io). Since there are > 1300 papers on this topic, we cannot read all of them😭, and we need your help on uncovering all the "hidden gems"!
Denoising Diffusion-based Generative Modeling: Foundations and Applications
From cvpr2023-tutorial-diffusion-models.github.io
56K
Jiaming Song
@baaadas
Jun 17, 2023
📢 Our #CVPR2023 tutorial on "Denoising Diffusion Models: A Generative Learning Big Bang" w/ @chenlin_meng and @ArashVahdat is happening tomorrow morning! 9:00 to 12:30, West 202-204. …3-tutorial-diffusion-models.github.io This is the year of big bang for diffusion models in CVPR!
41K
Jiaming Song
@baaadas
Jul 12, 2023
After a wonderful year at NVIDIA, I am starting a new adventure @LumaLabsAI 🐻
00:00
49K
Jiaming Song
@baaadas
Jun 12, 2024
Extremely proud to be working on this with many amazing people @LumaLabsAI! Generate a 5-second, 120 frames video in 120 seconds from text or images *now* on: lumalabs.ai/dream-machine Available to everyone. #LumaDreamMachine
26K
Jiaming Song
@baaadas
Jun 15, 2021
Introducing Diffusion-Denoising Models with Contrastive Representations (D2C), a non-adversarial image generative model for few-shot conditional generation (e.g. image manipulation). d2c-model.github.io arxiv.org/abs/2106.06819 w/ @a7b2_3 @chenlin_meng @StefanoErmon
Jiaming Song
@baaadas
Mar 14, 2025
Came across this gem earlier. TLDR: perplexity is a flawed metric for diffusion language models due to model mis-specification, so other metrics like the "Sequence Error Rate" proposed here might be better.
15K
Jiaming Song
@baaadas
Jul 13, 2025
Based on developments on "flow-map / average velocity" type methods, such as consistency trajectory models, shortcut models, IMM, and mean flow, I believe that the community will develop a proper replacement to diffusion / flow matching in 6 - 12 months.
Jiaming Song
@baaadas
Mar 11, 2025
As one of the people who popularized the field of diffusion models, I am excited to share something that might be the “beginning of the end” of it. IMM has a single stable training stage, a single objective, and a single network — all are what make diffusion so popular today.
20K
Jiaming Song
@baaadas
Dec 2, 2017
Basically you can use all kinds of regularization to maximize MI between data and code - e.g. GAN, Stein and MMD. Our experiments on PixelCNN show that MMD works the best, and can be implemented in 10 lines.
Stefano Ermon
@StefanoErmon
Dec 2, 2017
Check out our new blog post by @shengjia_zhao on InfoVAE: ermongroup.github.io/blog/a-tutoria…
Jiaming Song
@baaadas
Dec 8, 2020
Can we make better use of negative samples in contrastive learning? In our #NeurIPS2020 paper, we show this is true by simply using a multi-label objective. Come to our oral presentation at 6:15 PT (neurips.cc/virtual/2020/p…) and poster at 9-11 for more details! @StefanoErmon