LibreVideoDepthAnythingb-depth

Video Depth Anything Base (ViT-B DINOv2 encoder + temporal DPT head, 113.1M params), repackaged for LibreYOLO's depth task. Relative inverse depth (higher = closer), temporally consistent across the frames of a video.

Non-commercial use only. These weights are CC BY-NC 4.0. The Small model (LibreVideoDepthAnythings-depth) is Apache-2.0 and has no commercial restriction. You are responsible for complying with the weight licence.

Source

Converted from depth-anything/Video-Depth-Anything-Base at revision 7231d0c6e260f54f103eba5005d01f6efa4f43db (video_depth_anything_vitb.pth). Code: DepthAnything/Video-Depth-Anything at commit 4f5ae23172ba60fd7bc11ef671cca678842c7072 (Apache-2.0). Copyright (2025) Bytedance Ltd. and/or its affiliates. The weights are licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

Chen, S., Guo, H., Zhu, S., Zhang, F., Huang, Z., Feng, J., and Kang, B. "Video Depth Anything: Consistent Depth Estimation for Super-Long Videos." CVPR 2025.

Modifications

None to the learned parameters: the upstream state dict is wrapped in LibreYOLO checkpoint metadata with unchanged keys and tensors. See weights/convert_video_depth_anything_weights.py in the LibreYOLO source repository.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreVideoDepthAnythingb-depth.pt")
results = model.predict("video.mp4")  # one Results per frame, one shared relative scale
depth = results[0].depth_map          # DepthMap (H, W), higher = closer

License

CC BY-NC 4.0. Commercial use is not permitted. See the LICENSE and NOTICE files in this repository.