Research Scientist @NVIDIA, Generative AI, reasoning across different data modalities, and compositionality with few or zero examples. Opinions are my own.
London, United Kingdom
Joined December 2016
- Tweets233
- Following206
- Followers588
- Likes783
Pinned Tweet
This was awesome work led by @NoamRot ! 🚀
We want generative models to go beyond pixels: 3D content we can edit, animate, and use in simulation.
JCA is a step toward this. It jointly generates video and 3D motion for any rig, with motion that responds to its surroundings.
🎬 Excited to share our paper: Joints, Camera, Action! (JCA)
JCA generates video & 3D motion together for any rig!
Video models generate motion you can watch. But what if you could use that motion to animate a 3D asset, with motion that responds to its surroundings?
research.nvidia.com/labs/par…
🧵
Yuval Atzmon retweeted
🚀 Introducing ReLift: dense 3D tracking from monocular video.
Strong 2D trackers provide correspondences.
Geometry models produce depth & cameras.
We show how to combine them, repair 3D errors, and recover occluded motion for SOTA 3D tracking.
All it takes: one small corrector + off-the-shelf models.
🌐 Project: relift3d.github.io/
@MarounKhriesh @NoamRot @yoni_kasten @orlitany
Yuval Atzmon retweeted
Coding agents can build an app. What if they could give it a body?
Meet Metatron: a versatile System 1 controller for humanoids. A goal-based interface connects System 2 planners, including coding agents, to whole-body movement.
A few examples below. 1/5
Yuval Atzmon retweeted
1/4 We trained an 8B safety classifier on just 1,275 carefully annotated prompts.
It beat GPT-5.4, Claude Sonnet 4.6, and leading open-weight guardrails trained on far more data, on average across 5 safety benchmarks.
Quality > quantity.
Yuval Atzmon retweeted
🎬 Excited to share our paper: Joints, Camera, Action! (JCA)
JCA generates video & 3D motion together for any rig!
Video models generate motion you can watch. But what if you could use that motion to animate a 3D asset, with motion that responds to its surroundings?
research.nvidia.com/labs/par…
🧵
Yuval Atzmon retweeted
Do vision-language models look at the image regions they describe? In our new paper, we present Self-Saliency, a method for aligning visual attention with reasoning text.
research.nvidia.com/labs/par… w/ @GalChechik @DalalGal
Yuval Atzmon retweeted
[1/4] EARLI is out in Nature's Communications AI & Computing!
EARLI combines learning and genetic algorithms to accelerate routing solvers like cuOpt and HGS, producing high-quality routes in seconds for hundreds of delivery locations.
Yuval Atzmon retweeted
Meet DyRAD: Radar NVS for dynamic driving scenes.
📡🚗 Radar directly senses motion through Doppler, but prior radar NVS methods ignore this signal where it matters most: dynamic scenes.
🌐 dyrad-nvs.github.io
📄 arxiv.org/abs/2609.39841
💻 github.com/Dyrad-NVS/DyRAD
🧵
Yuval Atzmon retweeted
🥳 Excited to share UniProbe: A learnable, token-level hallucination detection for large VLMs 🔎
UniProbe looks inside a VLM’s internals to identify hallucinated response tokens and mitigate them before they reach the user.
📉 55% fewer hallucinations
⚡ Just 1.06× latency
Yuval Atzmon retweeted
I'm going to present ByG at #ICML26 today (Thursday) @ 2:30pm - poster #2615!
Stop by if you're interested in unsupervised editing with flow matching!
Project page: research.nvidia.com/labs/par…
Details below ⬇️
🎉 Happy to share our latest work: Bootstrap Your Generator: Unpaired Visual Editing with Flow Matching (accepted to #ICML2026)!
TL;DR: We train image and video editing models without any paired data.
No ground-truth edit data, no external reward models.
Yuval Atzmon retweeted
I'm going to present Fast-AR at #ICML26 tomorrow (Tuesday) @ 2pm - poster #1204!
Come and say hi 👋
If you're around this week, feel free to DM me.
Project page: dvirsamuel.github.io/fast-au…
Details below ⬇️
Yuval Atzmon retweeted
That's why we built github.com/NVlabs/ProtoMotio…
Reproducible results, open source, a shared home for robotics and animation.
(Currently focused on humanoids, but can be extended to fixed-based/quads).
This was fun work! A cool part of ByG is making cycle consistency work for denoising models. The missing piece was a Straight-Through trick. ST usually trains quantization or LLM sampling; here, ST lets reverse pass see a clean generated video while grads flow through a surrogate
🎉 Happy to share our latest work: Bootstrap Your Generator: Unpaired Visual Editing with Flow Matching (accepted to #ICML2026)!
TL;DR: We train image and video editing models without any paired data.
No ground-truth edit data, no external reward models.
Yuval Atzmon retweeted
🎉 Happy to share our latest work: Bootstrap Your Generator: Unpaired Visual Editing with Flow Matching (accepted to #ICML2026)!
TL;DR: We train image and video editing models without any paired data.
No ground-truth edit data, no external reward models.
Yuval Atzmon retweeted
"Fast 4D Mesh Generation by Spatio-Temporal Attention Chains"
TL;DR: training-free spatio-temporal attention chains generate topology-consistent 4D meshes from video 13× faster while improving temporal correspondence quality
This was a fun collaboration! It's great to work with Dvir again
Yuval Atzmon retweeted
New blogpost out 📃
"Detecting LLM Misbehaviors from the Inside Out with Deep Learning on Structured Data" (ffabffrasca.substack.com/p/d…) [1/8]
This was a fun work, led by the skilled @sapiryiflach
T2I models struggle with spatial reasoning - "a dog to the right of a cat" often comes out wrong. Instead of handcrafting a test-time loss func, we trained a classifier on attention maps and used it as a differentiable loss
🚀Excited to present our new paper that has been accepted to #WACV2026!
Text-to-image models often fail at simple spatial tasks, like placing a dog to the right of a teddy bear.
Our solution: Learn-to-Steer.
We learn a loss function directly from attention maps and apply it during inference.
This work was done together with @AtzmonYuval and @GalChechik
📰arXiv: arxiv.org/abs/2509.02295
🌐Project page: learn-to-steer-paper.github.…
📽️Video: youtu.be/KaxRwlE-UFg
The cool thing - despite training on single spatial relations, Learn-to-Steer generalizes to multiple relations in a single image - even 5 objects with 3 relations. And it works across different diffusion architectures including MMDiT
