Co-founder @genmoai. Co-created denoising diffusion (DDPM), DreamFusion, Dream Fields. Ex Ph.D. @berkeley_ai, @googleai, @facebookai, @nvidiaai, @mit
San Francisco, CA
Joined July 2009
- Tweets660
- Following5.6K
- Followers8K
- Likes5.2K
Ajay Jain retweeted
the most predictive trait for successful research these days seems to be excessive carefulness, bordering on paranoia. so easy to make bugs, so hard to find them
agents, so far, mostly make this more difficult
Spotted in SF: a bus half covered with a Vanta ad and half covered with a Delve ad. Fight fire with fire.
Ajay Jain retweeted
University of California faculty and alumni won five Nobel Prizes this week, setting a new record for the most faculty of one institution to achieve this great honor in a single year 🏅💙 💛
These remarkable achievement highlight the ongoing contributions of America’s #1 public research university and the central role of federal funding in advancing world-changing scientific inquiry.
bit.ly/4qn6CZV
Ajay Jain retweeted
I'm really excited to share our new formulation for post-training diffusion models! Here's why I think this formulation has the potential to be quite useful 👇
I'm at ICML in Vancouver this week. We've been rapidly growing our team at Genmo across research and infrastructure. Let me know if you're interested in catching up. It's going to be an exciting year!
Photo from last week's team offsite.
Full 3D self-attention is the single most expensive operation in video diffusion transformers like Mochi, accounting for 50-70% of the total inference cost.
New work: We dig deep into the attention inside these models, revealing beautiful structure and sparsity.
Ajay Jain retweeted
Preprint of today: Wen et al., "Analysis of Attention in Video Diffusion Transformers" -- arxiv.org/abs/2504.10317
Super interesting insights on Video ViTs. Attention sinks, sparse attention (which can be properly sparsified), specific layers being more important, and more!
Very cool research on Mochi-1. Image-to-video, keyframes, and long video are here!
Never seen a R1 moment in video diffusion models??😰Can't things just emerge using very low cost??🧐Certainly can!!!!
🚀 Introducing Pusa now!
Pusa: Thousands Timesteps Video Diffusion Model — A single model that unlocks:
Text-to-Video →
• Image-to-Video
• Start/End Frames to Video
• Video Transitions
• Video Extensions
• Next-frame prediction
• Novel sampling algorithms(frame-independent noise)
• ...
Also note that the model still can do text-to-video generation!!
All for just ~$0.1k training cost (about 100 H800 GPU hours)!
🔍 Key Innovations:
- Frame-level noise control (FVDM-inspired) → Unmatched flexibility
- Non-destructive adaptation → Preserves base T2V capabilities
- Universal methodology → Applicable to other SOTA models (e.g., Wan2.1, Hunyuan Video)
💡 100% Open-source—code, data, and training scripts release today!
📽 Demos and details below ↓
#AI #VideoGen #OpenSource
1. What is it?
2. Why do I maybe need to change settings?
3. If so, what are the settings?
4. The settings button at the top does not do anything
Iteration velocity is key to great work. Recently reminded of one of my 2020 papers. Our experimental logbook had 166 pages by the end: meticulously full of plots, results, discussions and evaluations. At least 9 / 10 experiments will fail, or you aren't trying enough things!
On detailed experimental logbooks: @deepakpathak taught me this trick. Very useful! Not just for wet labs.
Pieter is a visionary in both startups and fundamental research. His lab is a talent magnet. There's a reason many of us have gone on to start large AI companies.
Personally, an NSF fellowship was an accelerant of ambition. It allowed me to tackle big, new problems and crazy ideas, before there was enough evidence to justify private interest.
And 6 years later, most of the resulting research has been applied widely in the United States!
Ajay Jain retweeted
The first Genmo offsite of 2025 ✅
We brought the team together for an incredible three-day retreat in Calistoga! From exploring a beautiful winery and enjoying amazing dinners to board games and relaxing spas, it was the perfect way to unwind and reflect on our goals for this upcoming year 😌
We’ve had a busy start to 2025—and we’re still hiring across all teams ✨
Join us 🥂
Had a lot of fun at our San Francisco office warming last week. Thanks to all who came!
Last week, we got to celebrate the opening of @genmoai first office in San Francisco. An incredible night celebrating them. You will be hard to find a team that works harder than they do, led by the great and powerful @parasjain and @ajayj_
