Depth-Anything-V2-Large — transformers format (original code)

Unofficial transformers port of Depth Anything V2 Large (depth-anything/Depth-Anything-V2-Large, depth_anything_v2_vitl.pth) that runs the original code. Loads with trust_remote_code. model.safetensors holds the checkpoint's tensors under their original names, converted from .pth (every tensor checked equal).

Not an official release. Depth Anything V2 was created by Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng and Hengshuang Zhao — https://github.com/DepthAnything/Depth-Anything-V2, paper "Depth Anything V2" (arXiv 2406.09414).

What it is

DepthAnythingV2ForDepthEstimation holds the official DepthAnythingV2's submodules (pretrained: DINOv2 ViT-L/14, depth_head: DPT) and computes depth with that class's own forward, returning the standard DepthEstimatorOutput (predicted_depth); the official infer_image and image2tensor are kept. Code vendored from DepthAnything/Depth-Anything-V2 @ a561b849 into dav2_dinov2.py / dav2_dpt.py (relative depth) and dav2_metric_dinov2.py / dav2_metric_dpt.py (metric_depth/); edits: imports, and Block → NestedTensorBlock where the original imported NestedTensorBlock as Block. DepthAnythingV2Processor (processing_depth_anything_v2.py, processor_config.json) wraps DepthAnythingV2ImageProcessor, the image2tensor transform.

transformers' own DepthAnythingForDepthEstimation (the -hf repos) is a re-implementation: on the same input it matches the original only for square inputs; for non-square inputs its DINOv2 position-embedding interpolation differs (0.8–1.9 % max relative depth difference measured on Small). This package runs the original modules instead.

Verified against the official code (same process, V100, 3 images incl. non-square): preprocessing, predicted_depth (also as a tuple), the hidden states, infer_image, the processor's post-processing and fp16 autocast outputs are bit-identical.

Usage

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

model = AutoModelForDepthEstimation.from_pretrained("Depth-Anything-V2-Large", trust_remote_code=True).eval().cuda()
processor = AutoProcessor.from_pretrained("Depth-Anything-V2-Large", trust_remote_code=True)
image = Image.open("image.jpg")                                   # or a path (read with cv2, as the original scripts)
inputs = processor(images=image, return_tensors="pt").to("cuda")
with torch.no_grad():
    outputs = model(**inputs)                                     # outputs.predicted_depth: (1, H', W') relative inverse depth
depth = processor.post_process_depth_estimation(outputs, target_sizes=[image.size[::-1]])[0]["predicted_depth"]
# or, exactly as the original scripts: model.infer_image(cv2.imread("image.jpg"))

• AutoModel loads the same class. output_hidden_states=True adds the DINOv2 features the DPT head reads; return_dict=False returns a tuple.
• post_process_depth_estimation resizes as the original infer_image (bilinear, align_corners=True), not with transformers' bicubic.

Runtime packages beyond torch / transformers: opencv-python, torchvision (as the original).

Licensing

• Weights (depth-anything/Depth-Anything-V2-Large): CC-BY-NC-4.0 (non-commercial) — LICENSE.
• Code (vendored files and the port): Apache-2.0 — LICENSE.code.
• Attribution, sources and the changes made here: NOTICE.

Citation

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

@article{depth_anything_v2,
  title={Depth Anything V2},
  author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Zhao, Zhen and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
  journal={arXiv:2406.09414},
  year={2024}
}

@inproceedings{depth_anything_v1,
  title={Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data}, 
  author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
  booktitle={CVPR},
  year={2024}
}