Depth Anything V2 Small β€” LiteRT, 518Γ—392, weight-only int8

A LiteRT (TensorFlow Lite) conversion of Depth Anything V2 Small (revision 5426e4f0f36572d16453bbda7a8389317b1bef99, Apache 2.0), by Depth Anything (Lihe Yang et al., Depth Anything V2, 2024). Weights are unchanged apart from weight-only int8 quantization.

  • Input: [1, 392, 518, 3] float32, RGB in [0, 1]. ImageNet normalisation is inside the graph. Portrait frames are transposed into the landscape input.
  • Output: [1, 392, 518] float32 relative disparity (bigger is nearer), raw scale. Dividing by the frame's maximum reproduces Apple's Core ML build (correlation β‰₯ 0.9999 on 21 test frames).
  • Converted with: litert-torch 0.10.0, torch 2.14.1, transformers 5.23.0, then ai-edge-quantizer 0.10.1 (weight_only_wi8_afp32). Runs on LiteRT 1.2.0 and later.

License

Apache 2.0, as the original Depth Anything V2 Small checkpoint. The Base, Large and Giant Depth Anything V2 models are CC-BY-NC-4.0 and are not used here.

Downloads last month
11
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Luken58/depth-anything-v2-small-litert

Finetuned
(10)
this model