Meta has released two new AI models that do something straightforward but genuinely useful: they let you point at objects in images and video and say exactly what you want to do with them.
SAM 3 understands detailed text descriptions. Tell it "people sitting down, but not wearing a red baseball cap" and it finds those exact people in your video. Traditional AI struggles with this kind of specificity — it gets confused by the gap between what you're saying and what's actually on screen. SAM 3 bridges that gap.
This matters because it makes video editing faster. Meta is building this into Edits, their video creation app, so creators can apply effects to specific people or objects without manually selecting them frame by frame. The same technology is coming to Meta AI and Vibes, their AI assistant tools.
We're a new kind of news feed.
Regular news is designed to drain you. We're a non-profit built to restore you. Every story we publish is scored for impact, progress, and hope.
Start Your News DetoxFrom flat to three-dimensional
SAM 3D does something more ambitious: it reconstructs 3D objects and entire scenes from a single photograph. Two separate models handle different tasks. One rebuilds objects and spaces. The other maps human bodies and shapes with precision — useful for fields like sports medicine and robotics, where you need accurate body data.
To measure whether these models actually work well, Meta collaborated with artists to build a new evaluation dataset. This sounds technical, but it matters: the dataset includes diverse, real-world images that challenge the models in ways generic test images don't. It's a more honest way to know if the technology holds up.
You can already see SAM 3D in action on Facebook Marketplace. The new "View in Room" feature lets you upload a photo of a couch or lamp, then see how it would look in your actual living room before you buy it. That's the kind of friction-reducing feature that changes whether someone actually makes a purchase.
Tools for anyone
Meta has released both models openly — meaning researchers and developers can download them, study them, and build on them. They've also created a playground where you don't need to code at all. Upload an image or video, type what you want, and watch the AI segment objects or transform a 2D photo into a 3D scene you can rotate and manipulate.
They're partnering with Roboflow, a platform for data annotation, so people can fine-tune these models for their own specific needs. Whether you're building something niche or just exploring what's possible, the barrier to entry just got lower.
The next step is seeing what creators and researchers actually build with these tools. Open-source AI models often surprise their makers — people find applications nobody anticipated. Right now, Meta has shown the obvious use cases: faster editing, better shopping experiences, robotics research. But there's usually more.
