🎉 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.

Thank you for explaining our work so wonderfully well!

It's amazing to see how far ProtoMotions has come since it's first release. If you are looking for a feature-rich and scalable framework that can train controllers on massive datasets, then checkout ProtoMotions!

At @nvidia, we built ProtoMotions to help us, and researchers world-wide, innovate quickly without compromising on applicability. We're proud to announce ProtoMotions3 -- our biggest release yet! 🧵👇

🎉 I am excited to present our new paper! Our paper improves personalization of text-to-image models, by adding one special cleaning step on top of existing personalized models. With just a single gradient update (~4 seconds on an NVIDIA H100 GPU) and a single image of the target concept, our method improves both text alignment and image alignment. For example, it improves LoRA by ( 7% / 14%). This is achieved by adding new loss terms and taking into account the prompt and seed. This work was done together with @dvir_samuel and @GalChechik. 🌐 Paper page: per-query-visual-concept-lea… 📄 arXiv paper: arxiv.org/abs/2508.09045 More details in the comments below.

What if video editing took seconds instead of hours? OmnimatteZero – by @dvir_samuel Prof. @GalChechik from @Bar_ilan, HUJI & OriginAI – Removes dynamic objects with their shadows and reflections, separates layers, and reinserts them in real time. 👉 arxiv.org/abs/2503.18033

How do generative models represent the notion of "an object"? In early 20th century, Gestalt psychologist studied this question for human perception. This paper now looks into related mechanisms in text-to-video models.

Personalizing image generation from a single image is still very hard. Check out this paper

This morning, I had the pleasure of attending #EMTech Europe 2025 in Athens, an international conference on emerging technologies. Sr. Director of @NVIDIA, Israeli @GalChechik , gave a fascinating talk on the future of #AI, moderated by @yanpal7 of @kathimerini_gr , highlighting its transformative impact across industries in our lives. Innovation which is defining our future.

Teach your text-to-image model to count

🎉 I'm happy to share that our paper, Make It Count, has been accepted to #CVPR2025! A huge thanks to my amazing collaborators - @YoadTewel, @SegevHilit , @hirscheran, @RoyiRassin, and @GalChechik! 🔗 Paper page: make-it-count-paper.github.i… Excited to share our key findings!

It's always a bit stressful: Testing a pre-trained model on new datasets, with a different population and setup. Gluformer predictions were beyond what we hoped for!

We have a new and revised GluFormer manuscript! We expanded our analyses considerably: now showing that our AI model for CGM can identify individuals at higher risk of declining glycemic control before it happens, and can predict long-term diabetes & cardiovascular mortality.

🚀 Excited to release the code and demo for ConsiStory, our #SIGGRAPH2024 paper! No fine-tuning needed — just fast, subject-consistent image generation! Check it out here 👇 Code: github.com/NVlabs/consistory Demo: build.nvidia.com/nvidia/cons…

MaskedMimic pre-trained model public release 🧑‍🎄 github.com/NVlabs/ProtoMotio… Some info in the thread on how to play with the model 1/

Interesting #ECCV2024 keynote on distribution shift. @sanmikoyejo discussed interpolation and extrapolation. There is a 3rd case: Composition. Interpolate for each component but extrapolate the combination. Can we do better with composition than worst-case extrapolation?

TL;DR - we improve text-to-image output quality by tuning an LLM to predict ComfyUI workflows tailored to each generation prompt Project page: comfygen-paper.github.io/ Paper: arxiv.org/abs/2410.01731 [1\4]

At inference, just give the LLM a new prompt a high score, and predict a prompt-specific flow. For more info, please read our paper or come find us at @eccvconf Work done with my amazing collaborators @adihaviv @yuvalalaluf, Amit Bermano, @DanielCohenOr1 and @GalChechik [4/4]

Thrilled to share that our paper, "Where's Waldo: Diffusion Features for Personalized Segmentation and Retrieval" has been accepted to NeurIPS 2024! 🎉 Paper: arxiv.org/abs/2405.18025 Project Page: dvirsamuel.github.io/pdm.git… #NeurIPS2024 @GalChechik @RamiBenAri1 @MatanLvy

Animators of human motion want to control some aspect of the movement, and not bother with others. MaskedMimic does exactly that using physics-based models. Check out the code and video