light434/da3metric-large-openvino

DA3METRIC-LARGE OpenVINO IR (FP32)

OpenVINO IR export of depth-anything/DA3METRIC-LARGE. Format conversion only. No training.

This model is an OpenVINO conversion of DA3-Metric-Large. It is not weight-quantized; the “Quantized” relationship is used only because Hugging Face currently does not provide a dedicated model-conversion/deployment-optimization relationship.

Static input 1 × 3 × 336 × 504 (float32 NCHW, RGB, ImageNet mean/std). Outputs: canonical depth_raw, sky. Metric depth: depth_m = focal * raw / 300.

Files

• da3metric-large.xml
• da3metric-large.bin (FP32, compress_to_fp16=False)

Usage

import openvino as ov

core = ov.Core()
compiled = core.compile_model("da3metric-large.xml", "GPU")  # or "CPU"
# pixel_values: float32 NCHW ImageNet-normalized, shape (1, 3, 336, 504)
depth_raw, sky = compiled([pixel_values])

Resize and ImageNet normalize before this call. They are not inside the IR.

License

Apache 2.0, same as depth-anything/DA3METRIC-LARGE.

@article{depth_anything_3,
  title={Depth Anything 3: Recovering the Visual Space from Any Views},
  author={Lin, Haotong and Chen, Sili and Liew, Junhao and Chen, Donny Y. and Li, Zhenyu and Shi, Guang and Feng, Jiashi and Kang, Bingyi},
  journal={arXiv preprint arXiv:2511.10647},
  year={2025}
}