Metric depth models

da2-metric-indoor and da2-metric-outdoor are the official Depth Anything V2 Metric (Small) checkpoints (depth-anything/Depth-Anything-V2-Metric-Indoor-Small-hf and ...-Outdoor-Small-hf), exported to ONNX with a dynamic input size and 8-bit dynamic quantisation for in-browser use with Transformers.js. The export matches the PyTorch output exactly before quantisation; quantisation changes depth by about 1-4% (median) on test photos.

They are used by the “Measure distance” tool on /3d-lab/.

Other models (copied unchanged from Hugging Face)

  • onnx-community/depth-anything-v2-small: Depth Anything V2 Small (Apache-2.0), used for the 3D relief. model_quantized.onnx runs on the CPU (WebAssembly), model_fp16.onnx on WebGPU.
  • Xenova/slimsam-77-uniform: SlimSAM, a compressed Segment Anything (Apache-2.0), used to cut out the object in “Full object” mode.

Each folder keeps its original README.md (model card and licence). The Transformers.js runtime and three.js are in /assets/lib/ with their licences, so the 3D Lab needs no outside services.