HyDen DA2 Relative Depth

HyDen is a hybrid dual-path vision encoder for high-resolution monocular geometry that pairs a low-resolution transformer branch for global context with a full-resolution CNN branch for fine detail, delivering state-of-the-art accuracy at a fraction of the inference cost of competing methods. This checkpoint is the HyDen variant of DepthAnything-v2-large.

Paper: Hyden: A Hybrid Dual-Path Encoder for Monocular Geometry of High-resolution Images (ICLR 2026). Code, loader, and inference: https://github.com/facebookresearch/metadepth

Model details

• Task: monocular relative depth estimation
• Output: unnormalized inverse depth
• Weights precision: FP32
• Checkpoint: hyden_da2_reldepth_vitl_fp32_f809396504.pth
• License: FAIR Noncommercial Research License

Usage

Loader, preprocessing, and inference live in the MetaDepth repository: https://github.com/facebookresearch/metadepth

Download the checkpoint directly from this page, or via the huggingface_hub API:

from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download(
    repo_id="facebook/hyden-da2-relative-depth",
    filename="hyden_da2_reldepth_vitl_fp32_f809396504.pth",
)

Citation

@inproceedings{zhang2026hyden,
  title     = {Hyden: A Hybrid Dual-Path Encoder for Monocular Geometry of High-resolution Images},
  author    = {Zaiwei Zhang and Marc Mapeke and Wei Ye and Rakesh Ranjan and JQ Huang},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year      = {2026},
  url       = {https://openreview.net/forum?id=2eL6yXLCh8}
}

Contact

Open an issue on facebookresearch/metadepth.