DA3MONO-LARGE — transformers format

Unofficial transformers port of Depth Anything 3 DA3MONO-LARGE (monocular model; relative monocular depth, sky mask). Loads with trust_remote_code; the depth_anything_3 package is not needed. model.safetensors holds the tensors of depth-anything/DA3MONO-LARGE, unchanged (re-saved by save_pretrained; every tensor checked equal).

Not an official release. Depth Anything 3 was created by Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Guang Shi, Jiashi Feng and Bingyi Kang (ByteDance Seed) — https://github.com/ByteDance-Seed/depth-anything-3, paper "Depth Anything 3: Recovering the Visual Space from Any Views" (arXiv 2511.10647).

What it is

DepthAnything3ForDepthEstimation.model is the package's network, built from the checkpoint's own object graph (config.json → da3_config, identical to the package's configs/da3mono-large.yaml) the way depth_anything_3.cfg.create_object builds it; forward runs the network's forward and returns its outputs as a DepthAnything3Output (ModelOutput; the network's depth is named predicted_depth, the other keys keep their names). The network code is the official code (ByteDance-Seed/depth-anything-3 @ 3d835ec1) vendored into flat files (da3_*.py); the only edits are imports and one renamed logger. DepthAnything3Processor (processing_depth_anything_3.py, processor_config.json) wraps DepthAnything3ImageProcessor, which runs the package's InputProcessor.

Verified against the official package (same process and seeds, V100, 4 images): preprocessing and every output of the forward are bit-identical (also as a tuple), in fp32 and under the package API's autocast.

Usage

import torch
from PIL import Image
from transformers import AutoModelForDepthEstimation, AutoProcessor

model = AutoModelForDepthEstimation.from_pretrained("DA3MONO-LARGE", trust_remote_code=True).eval().cuda()
processor = AutoProcessor.from_pretrained("DA3MONO-LARGE", trust_remote_code=True)
inputs = processor(images=[Image.open(p) for p in paths], return_tensors="pt").to("cuda")  # pixel_values (1, N, 3, H, W)
dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16  # the package API's autocast
with torch.no_grad(), torch.autocast("cuda", dtype=dtype):
    outputs = model(**inputs)  # outputs.predicted_depth (1, N, H, W), outputs.depth_conf, ...

• AutoModel loads the same class; return_dict=False returns a tuple. Depth is at the processed resolution, as the package API returns it.
• Backbone features: export_feat_layers=[i, ...] returns them in outputs.aux.
• Some outputs use random sampling as in the original (sky handling samples depths): seed torch for reproducible runs.
• Runtime packages beyond torch / transformers: addict, omegaconf, einops, opencv-python (as the original).

Licensing

• Code and weights: Apache-2.0 — LICENSE (ByteDance-Seed/depth-anything-3), as published on depth-anything/DA3MONO-LARGE.
• Attribution, sources and the changes made here: NOTICE.

Citation

If you use this model, please cite the original work, as its authors ask:

@article{depthanything3,
  title={Depth Anything 3: Recovering the visual space from any views},
  author={Haotong Lin and Sili Chen and Jun Hao Liew and Donny Y. Chen and Zhenyu Li and Guang Shi and Jiashi Feng and Bingyi Kang},
  journal={arXiv preprint arXiv:2511.10647},
  year={2025}
}