README.md

ONNX Export for Depth Anything V3

This directory contains the export script to convert Depth Anything V3 models to ONNX format.

Prerequisites

1. Download the official model from Hugging Face:

   • DA3METRIC-LARGE

2. Set up the Depth Anything 3 environment following the official repository:

   • https://github.com/ByteDance-Seed/depth-anything-3

Export Steps

1. Copy the export script

Copy export.py from this directory to the Depth Anything 3 repository:

cp export.py /path/to/depth-anything-3/src/depth_anything_3/

2. Modify the API for ONNX compatibility

Edit src/depth_anything_3/api.py and change the autocast dtype from bfloat16 to float16 for ONNX compatibility:

Before:

# Determine optimal autocast dtype
autocast_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
with torch.no_grad():
    with torch.autocast(device_type=image.device.type, dtype=autocast_dtype):
        return self.model(
            image, extrinsics, intrinsics, export_feat_layers, infer_gs, use_ray_pose, ref_view_strategy
        )

After:

# Determine optimal autocast dtype
autocast_dtype = torch.float16
with torch.no_grad():
    with torch.autocast(device_type=image.device.type, dtype=autocast_dtype):
        return self.model(
            image, extrinsics, intrinsics, export_feat_layers, infer_gs, use_ray_pose, ref_view_strategy
        )

Note: This change is required because ONNX does not fully support bfloat16 operations.

3. Run the export

cd /path/to/depth-anything-3/src/depth_anything_3/

python export.py \
    --model-dir DA3METRIC-LARGE \
    --height 280 \
    --width 504 \
    --output-dir /path/to/output/

This will generate DA3METRIC-LARGE.onnx in the specified output directory.

Export Options

Argument	Default	Description
--model-dir	DA3METRIC-LARGE	Path to the model checkpoint directory
--height	280	Input height (must be divisible by 14)
--width	504	Input width (must be divisible by 14)
--batch-size	1	Batch size for export
--opset	20	ONNX opset version
--device	cuda	Device for export (cuda or cpu)
--output-dir	.	Output directory for ONNX file
--demo-image	-	Optional image path to run inference demo

Model Outputs

The exported ONNX model has two outputs:

• depth: Metric depth map [B, 1, H, W]
• sky: Sky classification logits [B, 1, H, W]