Neogenics1987/distill-any-depth-small-onnx

Distill-Any-Depth Small (ONNX)

An ONNX conversion of Distill-Any-Depth Small (24.8M parameters, Depth Anything V2 Small architecture) by Xiankang He, Dongyan Guo, Hongji Li, Ruibo Li, Ying Cui and Chi Zhang (paper, code, original weights).

Changes made: the original small/model.safetensors weights were loaded into the transformers DepthAnythingForDepthEstimation class and exported to ONNX (opset 18, float32) with dynamic height and width (multiples of 14). The network and weights are otherwise unchanged; the output matches the PyTorch model to within 1e-5.

• Input pixel_values: [1, 3, H, W], float32, ImageNet mean [0.485, 0.456, 0.406] and std [0.229, 0.224, 0.225], H and W multiples of 14.
• Output predicted_depth: [1, H, W], relative depth where larger = nearer.

Licence

The upstream project states the MIT licence for its code (© 2025 AGI Lab, Westlake University) and its Hugging Face model cards declare Apache-2.0 (xingyang1/Distill-Any-Depth) and MIT (xingyang1/Distill-Any-Depth-Small-hf). This conversion is distributed under the same permissive terms; the notices of both must be kept. Please also cite the paper (arXiv:2502.19204).