# Video Depth Anything
Video Depth Anything, by ByteDance, puts a temporal head on the Depth Anything V2 encoder so depth stays consistent from frame to frame. LibreYOLO runs a list of frames, a folder, a video or a stream as one sequence and returns one depth map per frame, all on one relative scale.
Tasks: Depth. Install: pip install libreyolo.
Verified against LibreYOLO v1.8.0.

## Install

Video Depth Anything needs no optional extra.

```bash
pip install libreyolo
```

## Predict

Weights download from the LibreYOLO Hugging Face mirrors on first use.

**Video**

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreVideoDepthAnythings-depth.pt")

# One result per frame, streamed with at most one window of lag.
for result in model("clip.mp4", stream=True):
    depth = result.depth_map.data      # (H, W) relative inverse depth
```

**One image**

```python
from libreyolo import LibreYOLO, SAMPLE_IMAGE

model = LibreYOLO("LibreVideoDepthAnythings-depth.pt")
result = model(SAMPLE_IMAGE, save=True)   # a one-frame sequence
print(result.depth_map.data.shape)
```

**CLI**

```bash
libreyolo predict model=LibreVideoDepthAnythings-depth.pt source=clip.mp4 save=True
```

Any multi-image source is one sequence: a list of frames, a directory sorted
by name, a video file or a stream. `N` frames give `N` results, each with a
`result.depth_map` on that frame's canvas. Inside, frames run in upstream's
overlapping 32-frame windows, and each window is scale and shift aligned to
the ones before it, so the whole sequence shares one relative scale. A single
image is a one-frame sequence; upstream pads it to a full window with copies
of itself.

The output follows LibreYOLO's shared depth contract: relative inverse depth,
where higher values are closer and the values have no metric unit. The input
size is upstream's default of 518 for every encoder; frames wider than 16:9
run proportionally smaller, as upstream does. See the
[depth estimation](/docs/tasks/depth-estimation) task page for the payload.

## Variants

Three encoder sizes, `s`, `b` and `l`, one checkpoint each. The licence
differs by size (see below), so `s` is the one to use in a commercial
product. Upstream's metric-depth checkpoints are not hosted, and the
converter refuses them.

Training is not offered: `train()` raises. Convert an upstream relative-depth
checkpoint with `weights/convert_video_depth_anything_weights.py` in the
library.

## Export

| Task | ONNX | TorchScript | ExecuTorch | TensorRT | OpenVINO | Paddle | MNN | RKNN | Axelera | ncnn | TFLite | CoreML | Core AI |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Depth |  |  |  |  |  |  |  |  |  |  |  |  |  |

A "yes" means the export is supported. An empty cell means the exporter refuses that combination.

Export is not supported. The network reads a 32-frame window and the windows
are stitched on the host, which the single-image depth export contract does
not describe, so `export()` raises.

## Checkpoints

| File | Input (px) | Task | Weights license |
| --- | --- | --- | --- |
| `LibreVideoDepthAnythings-depth.pt` |  | Depth | apache-2.0 |
| `LibreVideoDepthAnythingb-depth.pt` |  | Depth | cc-by-nc-4.0 |
| `LibreVideoDepthAnythingl-depth.pt` |  | Depth | cc-by-nc-4.0 |

## Licensing

Check the license on the Hugging Face repository of the specific weights you download. That repository is authoritative and licenses are not always uniform across a family. This is a description of the licenses involved, not legal advice.

- Original work: Video Depth Anything, ByteDance
- Upstream license: Apache-2.0 (Small checkpoint); CC-BY-NC-4.0 (Base and Large checkpoints)
- Upstream source: https://github.com/DepthAnything/Video-Depth-Anything
- LibreYOLO code: Apache-2.0
- Weights: Apache-2.0 (Small checkpoint); CC-BY-NC-4.0 (Base and Large checkpoints), republished at https://huggingface.co/LibreYOLO
- Interpretation: The licence depends on the size. The s checkpoint (Video-Depth-Anything-Small) is Apache-2.0, a permissive license that allows commercial and closed-source use if you keep the license text and notices. The b and l checkpoints (Base and Large) are CC BY-NC 4.0: attribution is required and commercial use is not allowed. The vendored code is Apache-2.0 from DepthAnything/Video-Depth-Anything at commit 4f5ae23172ba60fd7bc11ef671cca678842c7072 (Copyright ByteDance), including a motion module from AnimateDiff and attention code from Hugging Face diffusers, both Apache-2.0; the DINOv2 encoder and DPT head are shared with LibreYOLO's Depth Anything V2 family, and LibreYOLO's own adapter and streaming stitcher are MIT (libreyolo/models/video_depth_anything/NOTICE). LibreYOLO mirrors the converted checkpoints, tensors unchanged, as LibreYOLO/LibreVideoDepthAnything{s,b,l}-depth with the licence of each upstream size, not gated; metric checkpoints are not hosted.

Video Depth Anything is by ByteDance
([DepthAnything/Video-Depth-Anything](https://github.com/DepthAnything/Video-Depth-Anything)).
The vendored code is Apache-2.0. The licence of the weights depends on the
size:

- `s` (Video-Depth-Anything-Small): Apache-2.0, commercial use allowed.
- `b` and `l` (Base and Large): CC BY-NC 4.0, **non-commercial only**.
  They ship with a non-commercial notice.

LibreYOLO mirrors the converted checkpoints, tensors unchanged, as
`LibreYOLO/LibreVideoDepthAnything{s,b,l}-depth`, not gated.

## Citation

```bibtex
@article{video_depth_anything,
  title={Video Depth Anything: Consistent Depth Estimation for Super-Long Videos},
  author={Chen, Sili and Guo, Hengkai and Zhu, Shengnan and Zhang, Feihu and Huang, Zilong and Feng, Jiashi and Kang, Bingyi},
  journal={arXiv:2501.12375},
  year={2025}
}
```

Copied from https://github.com/DepthAnything/Video-Depth-Anything#citation
