Or Hirschorn (@Or_Hirsch) - โปรไฟล์ Twitter

Or Hirschorn รีทวีตแล้ว

Inbar Huberman-Spiegelglas@inbarhub·4 Mar

🎉Splatent has been accepted to #CVPR2026 Get sharp, high-quality reconstructions from diffusion latent space. Research by Amazon Prime Video: pape:r arxiv.org/abs/2512.09923 project page: orhir.github.io/Splatent/ Huge thanks to the amazing team @Or_Hirsch @FritzLior @omeriko_

Or Hirschorn@Or_Hirsch

Happy to share Splatent, a new research done during my internship at Amazon @PrimeVideo! 🎬 We tackle a key issue in 3D generation: getting sharp reconstructions directly from diffusion latent space. 📄 Paper: arxiv.org/abs/2512.09923 🌐 Page: orhir.github.io/Splatent/

English

0

3

14

769

Or Hirschorn@Or_Hirsch·22 Şub

Splatent is accepted to #CVPR2026!! 🎉🚀 Huge thanks to the incredible collaborators who made this happen: @FritzLior, @inbarhub, and the rest of the team.

Or Hirschorn@Or_Hirsch

Happy to share Splatent, a new research done during my internship at Amazon @PrimeVideo! 🎬 We tackle a key issue in 3D generation: getting sharp reconstructions directly from diffusion latent space. 📄 Paper: arxiv.org/abs/2512.09923 🌐 Page: orhir.github.io/Splatent/

English

1

8

43

4.3K

Or Hirschorn รีทวีตแล้ว

Nimrod Shabtay@NimrodShabtay·15 Oca

Excited to share CLIMP - the first fully Mamba-based contrastive vision-language model. Unlike CLIP's ViTs, Mamba's state-space formulation favors locality & smoothness—better retrieval and OOD robustness. with @ItamarZimerman, @Eli_Schwartz and @RGiryes arxiv.org/abs/2601.06891

English

1

12

14

676

Or Hirschorn รีทวีตแล้ว

Kwang Moo Yi@kwangmoo_yi·18 Ara

Segre and Hirschorn et al., "Multi-View Foundation Models" Train an adapter to 2D foundational models like DINO/SAM/CLIP that allows turning them into "multi-view" versions. More reliable results when doing multi-view tasks.

English

3

10

89

6.7K

Or Hirschorn@Or_Hirsch·18 Ara

We believe this is a key step for robust 3D understanding and cross-view reasoning. Check out the code and visuals on our project page! Co-authored with @LeoSegre and Shai Avidan. #ComputerVision #FoundationModels #3D #DINO #SAM #CLIP

English

0

72

Or Hirschorn@Or_Hirsch·18 Ara

What can you do with it? 1️⃣ Robust Matching: Features lock onto 3D points across drastic view changes. 2️⃣ Segmentation: Click once, and MV-SAM segments across all views. 3️⃣ Geometry: Estimate globally consistent surface normals directly from features.

English

1

0

62

Or Hirschorn@Or_Hirsch·18 Ara

Excited to share our new paper: Multi-View Foundation Models! 🌐 We introduce a general framework to upgrade 2D foundation models (DINO, SAM, CLIP) into multi-view consistent foundational models. 📄 Paper: arxiv.org/abs/2512.15708 ⚡️ Project Page + Code: leosegre.github.io/Multi-View-Fou…

English

1

2

149

Or Hirschorn@Or_Hirsch·11 Ara

We achieve SOTA results on sparse and dense benchmarks! 🏆 Big thanks to my team for the guidance. We hope this aids future efficient 3D pipelines. #3DGS #GenerativeAI

English

0

1

92

Or Hirschorn@Or_Hirsch·11 Ara

💡 The Solution: Instead of forcing perfect 3D geometry, Splatent uses 3DGS for structure but offloads fine detail recovery to a 2D multi-view refinement step. By fixing artifacts in 2D renders rather than 3D space, we bypass VAE inconsistencies. 🧠

English

1

0

1

143

Or Hirschorn@Or_Hirsch·11 Ara

Happy to share Splatent, a new research done during my internship at Amazon @PrimeVideo! 🎬 We tackle a key issue in 3D generation: getting sharp reconstructions directly from diffusion latent space. 📄 Paper: arxiv.org/abs/2512.09923 🌐 Page: orhir.github.io/Splatent/

English

1

4

11

5.4K

Or Hirschorn รีทวีตแล้ว

Guy Yariv@guy_yariv·7 Oca

[1/8] Recent work has shown impressive Image-to-Video (I2V) generation results. However, accurately articulating multiple interacting objects and complex motions remains challenging. In our new work, we take a step toward addressing this challenge.

English

7

26

80

9.2K

Or Hirschorn@Or_Hirsch·12 Ara

🚀 Excited to announce that our demo is now live on @huggingface! 👉 Check it out here: huggingface.co/spaces/orhir/E… Give it a try and let me know your thoughts!

Or Hirschorn@Or_Hirsch

🚨 Excited to share our latest work: "Edge Weight Prediction For Category-Agnostic Pose Estimation". We propose a novel approach to CAPE, leveraging predicted weighted graphs for better keypoint localization. Paper: huggingface.co/papers/2411.16… Project Page: orhir.github.io/edge_cape

English

0

1

7

9.1K

Or Hirschorn@Or_Hirsch·26 Kas

4/ We hope EdgeCape inspires new ideas in category-agnostic pose estimation research. 🙌 Feel free to reach out if you have questions or are interested in collaborations! I'd love to hear your thoughts.

English

0

66

Or Hirschorn@Or_Hirsch·26 Kas

3/ 📊 Results: EdgeCape outperforms state-of-the-art methods on the MP-100 benchmark, achieving: ✅New SOTA in 1-shot settings ✅Superior performance among similar-sized methods in 5-shot settings. We also shine in cross-category and occlusion scenarios! 💪

English

1

0

94

Or Hirschorn@Or_Hirsch·26 Kas

🚨 Excited to share our latest work: "Edge Weight Prediction For Category-Agnostic Pose Estimation". We propose a novel approach to CAPE, leveraging predicted weighted graphs for better keypoint localization. Paper: huggingface.co/papers/2411.16… Project Page: orhir.github.io/edge_cape

English

1

0

2

9.2K

Or Hirschorn

ค้นพบ