# Depth Anything 3 (DA3-BASE) — card for AI agents

Open-weights vision transformer that predicts depth maps, camera poses and 3D geometry from one or more images.

Category: AI 3D model generators · maker ByteDance Seed Team · checked 2026-11-10

## Price
- Price: Free
- Free tier: Free open model weights under Apache 2.0, run locally

## License
Apache 2.0 (DA3-SMALL, DA3-BASE, metric and mono models); CC BY-NC 4.0 (DA3 Large and Giant) — https://huggingface.co/depth-anything/DA3-BASE

- commercial: **conditions** (Commercial use with conditions)
  > | [DA3-LARGE-1.1](https://huggingface.co/depth-anything/DA3-LARGE-1.1)                     | 0.35B     | ✅             | ✅            | ✅             |       |               |           | CC BY-NC 4.0     |
  (https://github.com/ByteDance-Seed/depth-anything-3/blob/main/README.md)
- attribution: **no** (No credit needed)
  > License Apache 2.0
  (https://huggingface.co/depth-anything/DA3-BASE)
- in app: **conditions** (Ship inside your app with conditions)
  > | [DA3-BASE](https://huggingface.co/depth-anything/DA3-BASE)                     | 0.12B     | ✅             | ✅            | ✅             |       |               |           | Apache 2.0     |
  (https://github.com/ByteDance-Seed/depth-anything-3/blob/main/README.md)

## Rules to follow
- Keep the Apache 2.0 license and notice files with any distributed copy of the model.
- Install and run the model locally from the GitHub repository; no hosted API is provided.
- Cite the Depth Anything 3 paper when used in research.
- The model may be weaker on domain-specific images; check results.
- DA3 Large and Giant checkpoints are CC BY-NC 4.0 (non-commercial); Small, Base, metric and mono models are Apache 2.0.

## How to connect
- Download only
- Repository: https://github.com/ByteDance-Seed/depth-anything-3
- Website: https://huggingface.co/depth-anything/DA3-BASE

## Notes
Added automatically; license read from the official pages.

## Numbers (AgentGid data)
- Gid Score: 29
- GitHub stars: 6,712

## Lines to paste into your prompt or AGENTS.md

```
- Depth Anything 3 (DA3-BASE) (AI 3D model generators). Open-weights vision transformer that predicts depth maps, camera poses and 3D geometry from one or more images. Price: free; commercial use with conditions; no credit needed; ship inside your app with conditions.
  Connect: download from https://huggingface.co/depth-anything/DA3-BASE (no API).
  Rule: Keep the Apache 2.0 license and notice files with any distributed copy of the model.
  Rule: Install and run the model locally from the GitHub repository; no hosted API is provided.
  Rule: Cite the Depth Anything 3 paper when used in research.
  Rule: The model may be weaker on domain-specific images; check results.
  Rule: DA3 Large and Giant checkpoints are CC BY-NC 4.0 (non-commercial); Small, Base, metric and mono models are Apache 2.0.
  License: https://huggingface.co/depth-anything/DA3-BASE (checked 2026-11-10; details: https://agentgid.com/assets/depth-anything/)
```

## Links
- AgentGid page: https://agentgid.com/assets/depth-anything/
- Is it free for commercial use? https://agentgid.com/assets/depth-anything/commercial-use/
- Source: https://huggingface.co/depth-anything/DA3-BASE


## Depth Anything 3: depth and 3D scenes from photos

Depth Anything 3 is a free AI model from ByteDance that works out how far away each part of a photo is. From several photos or video frames of a place, it also finds where the camera stood and builds a 3D scene. This card is about the Base size, which is under Apache 2.0; most larger sizes are non-commercial.

### Where to use it
- A 3D scene from a few photos: Take several photos of a room or an object and get one 3D scene file that a viewer or Blender can open.
- Where the camera stood: Work out the camera position and lens of each photo in a set, for example to line up photos or add 3D objects into them.
- Depth from a video: Run a phone video through the command line tool to get depth for its frames and a 3D view of the scene.
- Plug into Blender or robots: Feed depth into tools people already built around it, such as a Blender add-on, image workflows, robots or a web VR viewer.
- Blurred backgrounds and photo depth effects: Use the depth map of one photo to blur the background behind a person or product, or to make a slight 3D motion effect. (idea, not from the official pages)
- Rough room measurements: Get a first idea of room sizes from photos for a planning app, using the metric model that gives distances in meters. (idea, not from the official pages)

### How to start
- Get the code from GitHub and install it in a Python environment with PyTorch 2.

```
git clone https://github.com/ByteDance-Seed/depth-anything-3
cd depth-anything-3
pip install -e .
```

- Load the Base model by its name and run it on your photos; the weights come from Hugging Face on first use.

```
model = DepthAnything3.from_pretrained("depth-anything/da3-base")
```

- Or use the da3 command on a folder of photos or a video and export a GLB scene.

```
da3 auto path/to/images \
    --export-format glb \
    --export-dir output \
    --model-dir depth-anything/da3-base
```


### License
The Base model is under Apache 2.0: free for paid products, keep the license and notice files with copies. The Giant, Large and Nested models are for non-commercial use only.

### Good to know
- The project's own main example uses the Nested Giant model, which is non-commercial. For a product, name the Base, Small, Metric or Mono model yourself.
- The license of the Large 1.1 model is unclear. Its model page says Apache 2.0, while the project's table says CC BY-NC 4.0, so treat it as non-commercial until that is fixed.
- Base gives relative depth (nearer or farther), not meters. For distances in meters, the project offers a separate Metric model.
- It learned only from public academic image sets and may do worse on special kinds of pictures, poor light or busy scenes.
- The 3D Gaussian export needs the Giant models, which are non-commercial, so Base cannot make it.

Checked 2026-11-10 against: https://huggingface.co/depth-anything/DA3-BASE, https://huggingface.co/depth-anything/DA3-BASE/raw/main/README.md, https://huggingface.co/api/models/depth-anything/DA3-BASE, https://huggingface.co/depth-anything/DA3-GIANT/raw/main/README.md, https://huggingface.co/depth-anything/DA3-LARGE/raw/main/README.md, https://huggingface.co/depth-anything/DA3-LARGE-1.1/raw/main/README.md, https://huggingface.co/depth-anything/DA3NESTED-GIANT-LARGE/raw/main/README.md, https://huggingface.co/depth-anything/DA3-SMALL/raw/main/README.md, https://huggingface.co/depth-anything/DA3METRIC-LARGE/raw/main/README.md, https://huggingface.co/depth-anything/DA3MONO-LARGE/raw/main/README.md, https://github.com/ByteDance-Seed/depth-anything-3/blob/main/README.md, https://github.com/ByteDance-Seed/depth-anything-3/blob/main/LICENSE, https://github.com/ByteDance-Seed/depth-anything-3/blob/main/pyproject.toml, https://github.com/ByteDance-Seed/depth-anything-3/blob/main/docs/API.md, https://github.com/ByteDance-Seed/depth-anything-3/blob/main/docs/CLI.md, https://depth-anything-3.github.io/

### Mistakes to avoid
- Don't: Copying the README example with depth-anything/DA3NESTED-GIANT-LARGE into a commercial product → Do: from_pretrained("depth-anything/da3-base") (or DA3-SMALL, DA3METRIC-LARGE, DA3MONO-LARGE, all Apache 2.0) (The Nested Giant model is CC BY-NC 4.0, non-commercial only.)
- Don't: Treating Base output as distances in meters → Do: Use DA3METRIC-LARGE and metric_depth = focal * net_output / 300 for meters (The main models give relative depth; the formula is for the Metric model.)
- Don't: Asking DA3-BASE for gs_ply or gs_video export → Do: Use glb, npz, ply or depth_vis with Base; Gaussian exports need infer_gs=True with a Giant model (non-commercial) (Gaussian output is only available on the Giant models.)
- Don't: Starting da3 backend with --host 0.0.0.0 on a shared or public machine → Do: Keep the default 127.0.0.1 binding unless the network is trusted (The backend server has no login, so anyone who can reach it can use it.)

### Questions
- Can I use it in a paid app? Yes with DA3-BASE, which is Apache 2.0. The Giant and Nested models are non-commercial only.
- Which files can it export? GLB scenes, NPZ and PLY data, depth pictures and more; formats can be combined with a hyphen, such as mini_npz-glb.
- Does it work with one photo? Yes. It predicts a depth map from a single photo, and gives more consistent results from several photos of the same place.
- Is there an online demo? Yes, the model page links a demo on Hugging Face; the model itself runs on your own machine.
- Which Python versions work? Python 3.9 to 3.13 for the package; the optional web app part needs Python 3.10 or newer.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
