Depth Anything 3: Turn Any Photo or Video into a Hyper-Accurate 3D World

🤯 The Visual Geometry Revolution is Here

Get ready to throw out your complex 3D scanners! A groundbreaking new model called Depth Anything 3 (DA3) is changing how we perceive and reconstruct the world from ordinary images and videos. This project is a massive leap forward in visual geometry, capable of recovering hyper-accurate 3D spatial information from literally any visual input—whether it's a single snapshot, a video clip, or multiple views from a car.

If you're fascinated by AI, computer vision, or 3D technology, you need to see this model in action.

---

What is Depth Anything 3?

Depth Anything 3 is the latest state-of-the-art model designed to predict spatially consistent geometry and metric depth from visual inputs, even when the camera's position isn't known.

The core technology is surprisingly simple, yet revolutionary:

The Secret: Instead of relying on complex, specialized architectures or multi-task learning, DA3 achieves its stunning results using just a single, plain transformer (like a vanilla DINOv2 encoder) trained on a novel depth-ray representation.

This elegant simplicity allows the model to generalize incredibly well, setting new records across all major visual geometry benchmarks. It recovers depth and 3D space with superior geometric accuracy compared to all prior models, including its impressive predecessor, Depth Anything 2 (DA2).

You can dive into the technical details and results on the official project page.

Key Abilities That Will Blow Your Mind

DA3 isn't just a research paper—it's a practical tool with real-world applications that are already impacting fields like robotics, virtual reality, and autonomous systems.

1. Any-View Geometry: It can predict accurate 3D geometry from any number of views, excelling equally in monocular (single-image) depth estimation and multi-view scenarios.

2. State-of-the-Art Reconstruction: It provides the foundation for next-generation systems, boosting performance in areas like:

   • SLAM for Large-Scale Scenes: Improving Simultaneous Localization and Mapping (SLAM) performance, even surpassing traditional methods like COLMAP in efficiency and drift reduction.

   • Autonomous Vehicle Perception: Generating stable and fusible depth maps from multiple vehicle cameras to enhance environmental understanding.

3. Feed-Forward 3D Gaussian Splatting (3DGS): One of the most exciting features is its ability to generate high-quality Novel Views. By using DA3's output, it can instantly create a 3D Gaussian Splatting (3DGS) representation, allowing you to fly through the reconstructed scene and render photorealistic views from any angle.

🚀 Try the Live Demo!

The best way to appreciate the power of Depth Anything 3 is to try it yourself! The team has provided an interactive demo hosted on Hugging Face Spaces.

Depth Anything 3 Hugging Face Demo

How to Use the Demo:

1. Upload: Drop a video or a collection of images (landscape orientation is preferred).

2. Reconstruct: Click the "Reconstruct" button.

3. Explore: The demo will generate and display:

   • Metric Point Clouds: The raw 3D data points.

   • Metric Depth Map: The estimated distance of objects from the camera.

   • Novel Views (3DGS): If you enable the "Infer 3D Gaussian Splatting" option, you can render realistic, new views of your scene.

4. Measure: You can even click two points on your original image and the system will attempt to compute the real-world distance between them!

Whether you’re a hobbyist, a researcher, or just curious about the future of 3D, spend some time experimenting with Depth Anything 3. It's a clear sign that sophisticated 3D reconstruction is rapidly becoming accessible to everyone.