README.md

---
license: apache-2.0
base_model: depth-anything/Depth-Anything-V2-Metric-Hypersim-Small
tags:
  - depth-estimation
  - depth2depth
---

# depth2depth model files

The model files the [depth2depth](https://github.com/jeff-hykin/depth2depth) crate builds into itself. They are
Depth Anything V2 Small, metric, Hypersim (indoor, max depth 20 m), converted from the official
`depth_anything_v2_metric_hypersim_vits.pth`:

- `dinov2_vits14.safetensors`, `da2_head_vits.safetensors`: the encoder and the DPT head, for candle
  (converted by `tools/convert_weights.py`; includes the unused ImageNet classifier head candle's dinov2 loads).
- `da2_metric_hypersim_vits.onnx`: an ONNX export taking any input size that is a multiple of 14, for TensorRT
  (made by the crate's `tools/export_onnx.py`). `da2_metric_hypersim_vits_364x448.onnx` is the older fixed-size one.
- Prebuilt fp16 TensorRT engines for a 448x560 input (dimos's default), so these GPUs skip the minutes-long engine build:
  `..._448x560_trt100700_sm87.engine` (TensorRT 10.7, Jetson AGX Orin, JetPack 6 / L4T R36.3) and
  `..._448x560_trt101601_sm120.engine` (TensorRT 10.16, RTX 50-series). Other TensorRT versions, GPUs or sizes reject
  them and the crate builds its own from the ONNX. (`..._364x448_trt100700_sm87.engine` is an older fixed-size one.)

The crate pins these files by sha256 in its `model.json`. Weights are Apache-2.0, as the upstream Small model.