Roni Itkin

Researcher at The Hebrew University of Jerusalem | 3D Vision and Computer Graphics

[Roni Itkin @ItkRoni]

We introduce 🌍GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens.🌍

Most feed-forward 3DGS methods still start from pixel, voxel, or dense view-aligned primitives.
We take a different route: align first, decode later. 🧵👇

[Omer Benishu @omerbenishu]

🚨 Excited to share our new paper: "PhyGenHOI: Physically-Aware 4D Generation of Dynamic Human-Object Interactions"! 🎉 We tackle generating photorealistic 4D interactions by deeply coupling generative human motion diffusion with physical simulation! 🧠���🧵👇(1/4)

[Hadar Davidson @HadarDavidson]

Excited to share Colored Noise Sampling (CNS)!🎉
Instead of injecting white noise, our SDE sampler exploits the inherent spectral bias of diffusion models. We dynamically color the injected noise to focus on frequencies where details are missing, substantially improving FID.🧵1/9

[Anpei Chen @AnpeiC]

We suggest going back to relative pose modeling,
enabling efficient and robust 3D reconstruction with low memory overhead — in both streaming and offline settings.

Project: https://t.co/Ji6Uipd5sL
Paper: https://t.co/bUzsHX3XdQ

[Guy Yariv @guy_yariv]

Our paper:
"LaMI: Augmenting Large Language Models via Late Multi-Image Fusion"

has been selected for an Oral Presentation at #ACL2026!

LaMI boosts LLM visual commonsense by generating complementary images from a text prompt and late-fusing their evidence into the prediction
🧵

[Michael Baltaxe @MichaelBaltaxe]

Happy to share that our work "RAD: Retrieval-Augmented Monocular Metric Depth Estimation for Underrepresented Classes" was accepted to #CVPR2026 (findings).

Check out the project page: https://t.co/HwIMDBPxXd

[Avital Shafran @AvitalShafran]

Very excited to share the first paper from my postdoc, led by the talented @JieZhang_ETH . This was an extremely fun project with a great group of people 🥸

[Guy Yariv @guy_yariv]

✨Happy to share that our paper DyPE has been accepted to #ICML2026!

[Xingyu Chen @RoverXingyu]

Wonderful collaboration with @ItkRoni @IssacharNoam @YehonatanKe @RoverXingyu @AnpeiC @BenaimSagie !

[Anpei Chen @AnpeiC]

GlobalSplat: Stop unprojecting, start decoding. 🛠️

We fuse all input views into a fixed set of Global Scene Tokens to build high-fidelity 3D assets without the pixel-wise redundancy.

✅ Higher quality
✅ Better spatial allocation

🔗 https://t.co/m6ROsAOck8

#3DGS

[Zhenjun Zhao @zhenjun_zhao]

GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens

@ItkRoni, @IssacharNoam, @YehonatanKe, @RoverXingyu, @AnpeiC, @BenaimSagie

tl;dr: all input views->a fixed number of latent scene tokens->decoder->explicit 3D Gaussians

https://t.co/g038S2pRsz

[Roni Itkin @ItkRoni]

Joint collaboration with:
@IssacharNoam @YehonatanKe @RoverXingyu @AnpeiC @BenaimSagie

📄 Paper: https://t.co/lP6r5FXWmh
🌐 Project: https://t.co/hvONwLkqsK

[Roni Itkin @ItkRoni]

We introduce 🌍GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens.🌍

Most feed-forward 3DGS methods still start from pixel, voxel, or dense view-aligned primitives.
We take a different route: align first, decode later. 🧵👇

[Roni Itkin @ItkRoni]

Strong efficiency-quality operating point:
24 Views on A100:
1.79 GB peak GPU memory
77.88 ms inference
3.8 MB on disk
With as few as 2–32K Gaussians, 🌍GlobalSplat🌍 has better PSNR on RE10K then feed-forward 3DGS methods that use hundreds of thousands to millions of Gaussians.