LibreDetAny3D

Byte-identical mirror of the official DetAny3D full checkpoint, for use with LibreYOLO's LibreDetAny3D runtime adapter.

⚠️ NON-COMMERCIAL WEIGHTS, AND AN IMPLIED REDISTRIBUTION BASIS

These weights are not covered by LibreYOLO's MIT license. Two separate restrictions apply, and both bind you, the downloader:

1. Non-commercial depth lineage. Upstream documents the depth branch as initialized from the UniDepth v2 checkpoint, which is licensed CC BY-NC 4.0 — non-commercial. CC BY-NC 4.0 permits redistribution, which is why this mirror exists, but it restricts use to non-commercial purposes. This repository is tagged cc-by-nc-4.0 as the strictest known term.
2. No declared checkpoint license. OpenDriveLab released this file from a Google Drive folder with no license metadata. The DetAny3D repository declares Apache-2.0, which covers the code. LibreYOLO's maintainer approved this mirror by treating that project-level statement as the redistribution basis for the parameters DetAny3D itself trained. This is LibreYOLO's disclosed interpretation, not a clarification from the DetAny3D authors.

Review the terms of every component below before use, and seek your own legal advice for commercial applications.

Source

• Publisher: OpenDriveLab/DetAny3D
• Original distribution: the project's Google Drive folder (detany3d_ckpts/detany3d.pth)
• Upstream code revision: 10e484b
• SHA-256: cd5f737ddbf3ceb64f969141672d04c0a3d785ad3b8a72668619a530840cd217
• Size: 4,334,195,618 bytes

Components inside this checkpoint

The checkpoint is a single full-model file; its 3,612 parameter tensors all sit under a sam. prefix, with sam.image_encoder.dino.* and sam.image_encoder.depth_head.* subtrees. Component lineage, per upstream documentation:

Component	Upstream	License
SAM image encoder	facebookresearch/segment-anything	Apache-2.0
DINO backbone blocks	DINOv2	Apache-2.0
Depth branch (depth_head)	UniDepth v2, restructured	CC BY-NC 4.0
DetAny3D 3D heads	OpenDriveLab/DetAny3D	Apache-2.0 (project-level; see banner)

The depth lineage is stated by the DetAny3D authors, not verified here by key matching — the depth parameters are restructured and do not carry UniDepth's original key names.

Modifications

None. Mirrored byte-for-byte with its original filename and serialization. No tensor values or metadata were changed. It is not converted to LibreYOLO's v1.0 checkpoint schema — the adapter loads the unchanged official format.

LibreYOLO usage

import numpy as np
from libreyolo import LibreDetAny3D

with LibreDetAny3D("detany3d.pth", runtime_path="/path/to/DetAny3D") as model:
    result = model.predict("image.jpg", text=["car", "person"])
    print(result.boxes3d.xyz)

DetAny3D is a sibling API, not part of the generic LibreYOLO(...) factory. It requires the separately installed upstream runtime (set runtime_path= or DETANY3D_PATH), which brings its own dependencies including SAM, DINOv2, UniDepth and xFormers. That runtime carries its own licenses; installing it is your responsibility and this mirror grants no rights to it.

Text prompts additionally require GroundingDINO Swin-B and its checkpoint, which are not mirrored here — get them from IDEA-Research/GroundingDINO (Apache-2.0). Box and point prompts need no extra checkpoint.

Accepts box, point and text prompts with estimated calibration. Training, validation, tracking and export are not implemented; use upstream evaluation for 3D accuracy.

Validation

LibreYOLO's adapter was verified against the upstream runtime on nine CPU parity cases using this checkpoint, plus real prediction from the built wheel. This is adapter parity, not an independent 3D accuracy benchmark. CUDA parity and benchmark accuracy were not verified.

License

See LICENSE for the DetAny3D Apache-2.0 code license, the full UniDepth CC BY-NC 4.0 text, and the checkpoint's separate terms; and NOTICE for provenance and mirroring details.