LibreDepthAnything3l-depth

DA3MONO-LARGE ViT-L relative-depth weights converted for LibreYOLO's depth task.

Source

The architecture is derived from ByteDance-Seed/Depth-Anything-3 at commit 41736238f5bced4debf3f2a12375d2466874866d. Copyright (c) 2025 ByteDance Ltd. and/or its affiliates. Licensed under the Apache License, Version 2.0. The DINOv2 backbone components retain Meta Platforms' Apache-2.0 attribution.

The learned parameters come from depth-anything/DA3MONO-LARGE at revision f465978e618db8cc79c83b8bbf24964857db1875, published under Apache-2.0.

Modifications

State-dict key remapping only: conversion removes the official high-level API's outer model. prefix and wraps all 406 unchanged tensors in LibreYOLO checkpoint schema v1.0. LibreYOLO reproduces the official mono sky handling and maps the model's positive relative depth to relative inverse depth, where higher means closer and no metric unit is implied.

See weights/convert_depth_anything3_weights.py in the LibreYOLO source repository.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreDepthAnything3l-depth.pt")
result = model("image.jpg")[0]
inverse_depth = result.depth_map.data

License

Apache License 2.0. See the LICENSE and NOTICE files in this repository.