DAP-weights-safetensors

This repository hosts a safetensors-format mirror of Insta360-Research/DAP-weights, the official checkpoint released with the paper Depth Any Panoramas (DAP): A Foundation Model for Panoramic Depth Estimation by Lin et al. (Insta360 Research Team, 2025).

The model is a metric depth estimator for 360° × 180° equirectangular panoramas: DINOv3-ViT-L/16 encoder + DPT depth head + a "range-mask" head that flags pixels beyond the model's confident distance range.

Nothing here is retrained. The weights are bit-identical to the upstream checkpoint; this mirror only changes the on-disk format (PyTorch .pth → .safetensors) and removes storage-shared alias keys from the state-dict to keep the file ~1.46 GB (matching the original) instead of ~2.5 GB.

Files

File	Size	Description
dap_vitl.safetensors	~1.46 GB	All trainable tensors. 498 unique tensors (the upstream .pth exposed 858 keys; 360 were storage-aliased duplicates from DINOv3Adapter.blocks ↔ DINOv3Adapter.model.blocks — see "Implementation notes" below).
LICENSE.dap.md	—	DAP MIT license (Insta360 Research Team's contribution).
LICENSE.dinov3.md	—	DINOv3 License Agreement — governs the DINOv3 backbone weights inside the checkpoint, which are derivative works of Meta's DINOv3.

License

The weights in this repository are governed by two licenses simultaneously:

1. DAP / Insta360 contribution → MIT (LICENSE.dap.md)
2. DINOv3 backbone weights → DINOv3 License Agreement

The DINOv3 License is permissive but not MIT. By using these weights you agree to its terms, including (non-exhaustive):

• You may use, reproduce, distribute, copy, modify, and create derivative works of the DINO Materials.
• Your use must comply with applicable laws and Trade Controls (US OFAC, UN, EU, UK sanctions; export controls).
• You may not use the model for, or encourage others to use it for, any activities subject to ITAR or end-uses prohibited by Trade Controls — including military or warfare purposes, nuclear industries or applications, espionage, or the development or use of guns or illegal weapons.
• If you redistribute the weights or any derivative, you must include a copy of the DINOv3 License Agreement alongside them.
• You may not reverse engineer, decompile, or discover the underlying components of the DINO Materials.

See LICENSE.dinov3.md for the full text.

Acknowledgements

• Upstream weights: Insta360-Research/DAP-weights
• DAP paper / code: Insta360-Research-Team/DAP on GitHub
• DINOv3 backbone: Meta AI Research (DINOv3 License Agreement applies to the encoder weights inside this checkpoint).
• DPT decoder lineage: Depth Anything V2.

Neither Meta, Insta360 Research Team, nor any of the upstream authors endorse or are affiliated with this mirror.

Citation

@article{lin2025dap,
  title  = {Depth Any Panoramas: A Foundation Model for Panoramic Depth Estimation},
  author = {Lin, Xin and Song, Meixi and Zhang, Dizhe and Lu, Wenxuan and Li, Haodong and Du, Bo and Yang, Ming-Hsuan and Nguyen, Truong and Qi, Lu},
  journal = {arXiv:2512.16913},
  year   = {2025}
}

Usage

The ergonomic path: install ComfyUI-DAP, which auto-downloads this file on first run.

Direct PyTorch load:

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

path = hf_hub_download("apozz/DAP-weights-safetensors", "dap_vitl.safetensors")
state = load_file(path)

# Load into a DAP model built with the upstream architecture.
# Use strict=False: storage-shared alias keys (DINOv3Adapter.blocks vs
# DINOv3Adapter.model.blocks) are deduplicated in this checkpoint and the
# missing aliases are populated automatically via the shared underlying storage.
model.load_state_dict(state, strict=False)

Implementation notes

The upstream model.pth was a flat state-dict (plus an epoch integer) where 360 tensor keys were storage-aliased duplicates: PyTorch's state_dict() walks both DINOv3Adapter.model.blocks.* (canonical) and DINOv3Adapter.blocks.* (alias registered by self.blocks = self.model.blocks in DINOv3Adapter). torch.save stored each unique storage once; safetensors has no storage-deduplication mechanism, so a naive .clone()-then-save_file would double the file size. This mirror keeps only the canonical .model.* key for each storage. When loaded into a fresh DAP model, the alias parameters are populated automatically because they share underlying storage with the canonical parameter.