We introduce ObjectDrop, our recent @GoogleAI project, aimed at achieving photorealistic object removal and insertion. Explore our project page: https://t.co/GOj5uAIF3v Arxiv: https://t.co/0tNxic4mUI

Excited to share this has now been accepted at #NeurIPS2025 as a position paper (<6% acceptance)!🎉 We advocate for systematically studying entire model populations via weight-space learning, and argue that this requires charting them in a Model Atlas. @NeurIPSConf #NeurIPS 🧵👇

We are releasing a paper I'm very excited about. We know test-time scaling is a path to greatly improved results, and achieves reasoning in the case of LLMs. We present a new and promising way to amortize it into training using HyperNetworks for image generation models.

@karpathy Thanks for the inspiring talk (as always!). I'm the author of the Model Atlas. I'm delighted you liked our work, seeing the figure in your slides felt like an "achievement unlocked"🙌Would really appreciate a link to our work in your slides/tweet https://t.co/rJjBhMmRjf

When we "erase" a concept from a diffusion model, is that knowledge truly gone? 🤔 We investigated, and the answer is often 'no'! Using simple probing techniques, the knowledge traces of the erased concept can be easily resurfaced 🔍 Here is what we learned 🧵👇

A quick look on what we were working on in the past months 🎉 more exciting things to come 🙌🏻

Google presents LightLab Controlling Light Sources in Images with Diffusion Models

In our #ICLR2025 paper, we introduce WIND 🌬️ A method that embeds a distortion watermark directly in the diffusion noise! Our method ensures that the watermark in one image does not reveal information about the watermark in other images 🤫 📝 https://t.co/jmEw4g3scN (1/5)

🚨 New paper alert! 🚨 Millions of neural networks now populate public repositories like Hugging Face 🤗, but most lack documentation. So, we decided to build an Atlas 🗺️ Project: https://t.co/1JpsC6dCeg Demo: https://t.co/4Xy7yLdIZY 🧵👇🏻 Here's what we found:

Crazy work by colleagues in my team!

🚨 Excited to share ObjectMate our latest from @GoogleAI for zero-shot subject-driven generation & insertion 🚨 🔗 project page:  https://t.co/uBeJgccH3W 📄Arxiv:  https://t.co/J2Ya8TLLeP

Many thanks to my great collaborators @ShulAsaf Matan Cohen @DanaBerman9 @Yxp52492 Alex Rav-Acha @YHoshen

5/5 For object insertion, we leveraged an ObjectDrop ( https://t.co/qqmKcCLsZP) data synthesis stage, where we remove objects with their shadows/reflections. This creates high-quality background images for supervised training, leading to SOTA results with either 1 or 3 references.

4/5 Our dataset powers supervised subject-driven generation. Given 3 reference images of an object and a text prompt, our model generates the object in a new context with remarkable identity preservation - all without test-time fine-tuning.

3/5 We also find that larger datasets have higher rates of repeating objects. From a web-scale dataset with 55M detected objects, we extracted 4.5M objects, each with at least 3 distinct views with diverse poses and scenes.

2/5 But finding these repetitions requires specialized similarity features. While others have used semantic features like CLIP or DINO, we show that it’s crucial to use tailored features for instance retrieval.

1/5 Our analysis reveals that large datasets like WebLI contain objects that reappear in different poses and scenes (e.g., car models, laptops, IKEA furniture). We call this the Object Recurrence Prior. We used it to create a dataset of 4.5M objects, each with multiple views.

🚀 Excited to share ObjectMate, our latest @GoogleAI project! A new approach to zero-shot subject-driven generation and object insertion. 🔗 Explore our project page: https://t.co/8ymD8QiiIT 📄 Arxiv: https://t.co/Mvdr4lY5Wk

I'm sharing something unique we've been making at Google (w/ UNC). We are releasing our work on a new class of interactive experiences that we call generative infinite games, essentially video games where the game mechanics and graphics are fully subsumed by generative models 🧵

Giving a talk about common neurons in vision models and emergent representations in diffusion model weights today at @eccvconf ☺️

With friends at @Google we announce 💜 Magic Insert 💜 - a generative AI method that allows you to drag-and-drop a subject into an image with a vastly different style achieving a style-harmonized and realistic insertion of the subject (Thread 🧵) web: https://t.co/32gsziKge1